Intelligence isn't an add-on: it's infrastructure. Embeddings run locally with Ollama, semantic search lives on PostgreSQL, and modules talk to models through MCP — the Model Context Protocol that exposes the platform as a set of safe tools.
The platform's data layer sits on PostgreSQL, with
pgvector for embeddings next to
the data. Semantic intelligence runs locally with Ollama; integration with models goes through a standard
protocol, MCP. More control, data that stays in-house, and synchronization keeping the underlying
operational systems aligned.
Identity, telemetry, documentation and embeddings in the same engine. Semantic search happens where the platform data already lives, with real referential integrity.
Embedding models run on the company server. Documents, notes and knowledge base become vectors without anything sensitive leaving the company perimeter.
Modules expose tools that models can invoke in a controlled way: list modules, search documents, scaffolding, archiving. Every call is explicit and traceable.
An MCP server (Node.js) operates on the whole project and exposes typed tools. It doesn't work "blind": it knows modules, forms, documents and registration rules. This is how AI acts on the platform with precision — both in development and at runtime. Generic AI starts from scratch; here it builds on a production-proven engine, on safe rails.
MCP server · stdio / HTTP · operates on the whole project
scaffold_module) inherits the
central nervous system — dashboard (DAH), telemetry (SEH), audit, three-pillar
documentation, MLPivot, log, Nginx and PostgreSQL synchronization — without writing them by hand.
See the cross-cutting infrastructure →
AI accesses the sources and the MCP tools: it reads the .dpk files, understands conventions, proposes and generates coherent code, scaffolds new modules and verifies registration. An order of magnitude faster than traditional development.
The production dashboard integrates AI-assisted analysis. Semantic search over documents (Ollama embeddings + pgvector) powers contextual help and support. AI works on real data, traced via telemetry.
Embeddings are generated by Ollama on the company server. No sensitive data leaves the perimeter to be vectorized.
The MCP server can run over stdio or HTTP with token and TLS. Tools have explicit contracts: you grant only what's needed.
Actions go through telemetry on PostgreSQL: who, when, what. The same audit infrastructure as the platform.
PostgreSQL for data, Ollama for embeddings, MCP for tools. In-house, traced, repeatable.